---
title: "Compute Em Level Endpoint"
method: POST
path: "/api/v1/cpt/em-level"
tags: ["cpt"]
---

# Compute Em Level Endpoint

`POST /api/v1/cpt/em-level`

## Request

- Base URL: `https://api.accurecord.io` (the document declares no server; this is the origin it was published from)
- URL: `https://api.accurecord.io/api/v1/cpt/em-level`
- Auth: none declared

Compute an E/M level deterministically, with a citeable worksheet.

An LLM reads the note and reports WHAT IT SAYS — how many problems, which data, what
management risk, how many documented minutes — each with a verbatim quote. The
arithmetic that turns those into 99213 rather than 99214 is a pure function and never
touches a model.

**The worksheet is the product.** Every element carries the rule that decided it and the
spans it rested on, and the losing path is reported too: "the time basis supports 99215
but the MDM only supports 99214" is a fact a coder needs, and a calculator that reports
only its winner destroys it.

⚠ **A quote that is not in the note is discarded**, not trusted. An observation with no
surviving quote does not count toward any element.

**Requires Pro or Enterprise tier, an executed BAA, and PHI activation.**

## Headers

- `x-api-key` string, nullable

## Request body

- EmLevelRequest
  - `family` string, required: E/M family key, e.g. 'office_established'. ⚠ Required rather than inferred: whether a patient is new or established, and whether the setting is office or hospital, are facts about the encounter that a note frequently does not state and that this service cannot know. Guessing it would change the code.
  - `chart_text` string, nullable: Visit note. ⚠ PHI — requires an executed BAA and PHI activation.
  - `observations` EmObservationsRequest: Pre-extracted observations, for callers who did their own reading. ⚠ Supplying these SKIPS the LLM entirely, which is the point: a caller with a structured EHR feed, or one replaying a stored extraction, gets a byte-identical worksheet with no model call and no PHI leaving the request.
    - `problems` object[]
    - `data` object[]
    - `risk` object[]
    - `total_time_minutes` integer, nullable
    - `total_time_evidence` string[]

## Response `200`

Successful Response

- object

## Other responses

- `400`: Malformed or out-of-bounds request — a batch over 50 codes, a file over the upload limit, an unparseable identifier, or a surface that is switched off in this environment (the `cpt_*_disabled` gates are 400, not 403). ⚠ Not interchangeable with 422: 400 is raised BY the handler, 422 by request validation.
- `401`: Missing, malformed, revoked or expired credential. Send the key as `X-API-Key` or `Authorization: Bearer`. `detail` is a string, except the MFA challenge on `/auth/login`, which is `{code: 'mfa_required', message}`.
- `403`: Authenticated but not permitted: the tier does not include the feature, the role is insufficient, or the free tier is paused. Gate-specific bodies use the `{code, message}` shape (`tier_required`, `insufficient_role`). ⚠ The free-tier pause is the ONE error in the service that does not use `detail` at all — its body is `{error: 'free_tier_paused', message}` at the top level. ⚠ Preconditions the caller CAN satisfy (no org, no BAA, PHI mode off) are 412, not 403 — see below.
- `404`: Resource, code, or org-scoped record not found. Note that a record belonging to ANOTHER organization is also a 404, never a 403 — existence is not disclosed across tenants.
- `412`: A precondition on the calling principal is unmet and the caller can fix it: no organization linked to the key (`org_required`), no signed BAA (`baa_required`), PHI mode not activated (`phi_not_activated`), reviewer mode off (`review_mode_disabled`), or the session was never second-factored (`mfa_required`). All use the `{code, message}` shape.
- `422`: Request body or query failed validation. FastAPI's own failures are a LIST of `{loc, msg, type}`; semantic refusals raised by the service are an object (the risk-adjustment endpoints use `{status, ...}`).
- `429`: Rate limit exceeded for this key's tier, or the coding surface is at capacity. Read `X-RateLimit-Reset-Minute` for the backoff rather than retrying immediately — an AI coding call is metered as a billable unit and a retry storm is charged.
- `500`: Unhandled server error. The body is `{detail: 'Internal server error', request_id: <id>}` and the same id is on the `X-Request-ID` response header.
- `503`: The coding brain cannot serve this request: the LLM circuit breaker is OPEN after repeated upstream failures, or the CPT surface is switched off / its reference data is unbuilt. ⚠ A breaker that opens MID-request does NOT surface here — the pipeline degrades rather than fails and says so in the 200 body's `degradations[]` (kind `llm_circuit_open`). Read that list: a degraded answer is shaped exactly like a healthy one.

---

[API](https://skmtc.dev/accurecord/apis/icd-10-coding-api.md) · [All operations](https://skmtc.dev/accurecord/apis/icd-10-coding-api/llms.txt) · [OpenAPI document](https://skmtc.dev/accurecord/apis/icd-10-coding-api/revisions/b1e12ce06ba5?raw)